This repository contains the official implementation of the CBF-SSM model presented in Structured Variational Inference in Unstable Gaussian Process State Space Models by Silvan Melchior, Felix Berkenkamp, Sebastian Curi, Andreas Krause.
Please cite the above paper when using this code in any way.
The datasets PR-SSM was already benchmarked on (Actuator, Ballbeam, Drive, Dryer, Furnace, Sarcos) can be downloaded as described in the readme in their repo.
The remaining datasets (RoboMove, Voliro, SpringNonLinear) can be downloaded here.
All datasets need to be placed in cbfssm/datasets/data.
To install CBF-SSM, run:
$ cd <path-of-repo>
$ pip3 install -e .
The folder run contains a script to reproduce the results for every dataset we use to compare CBF-SSM to previous work. The results will be in a new folder called run_output.
Follow these instructions to run your own experiments using CBF-SSM
At first, write a new dataset class which derives from the
base class. The code needs to overload dim_u, dim_y
and the method prepare_data (see example) s.t. it
- loads the data
- normalizes the data
- saves the data as train- and test-arrays with shape
[experiments, time-samples, data-dimension] - calls
create_batches()
Loading of the data depends on the source of your new dataset. For normalizing the data, there are helper functions if you have one experiment only (i.e. one long sequence), again see example.
Then, write a new run-file. You can use the template as a starting point, which also contains a lot of comments on how to choose your parameters.